Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces PhaseAT, a novel adversarial training framework for medical image domain generalization that leverages the semantic information encoded in Fourier phase. By generating phase-perturbed training views while preserving amplitude spectra, PhaseAT forces models to learn domain-invariant spatial structures, mitigating reliance on spurious texture correlations. The method achieves over 20% improvement in single-source DG performance, surpassing existing state-of-the-art approaches on challenging medical datasets.
Adversarial training on Fourier phase perturbations unlocks over 20% domain generalization gains in medical imaging by decoupling spatial structure from domain-specific textures.
Reliable clinical deployment of deep medical image models is hindered by distribution shifts across scanners, sites, and acquisition protocols. Existing domain generalization (DG) methods often focus on style or intensity diversification, but they can still leave networks dependent on domain-specific texture correlations. Inspired by evidence that Fourier phase encodes semantic structure, we introduce PhaseAT, a phase-aware adversarial training framework for medical DG. PhaseAT forms phase-perturbed training views in the Fourier domain by iteratively updating a bounded phase perturbation while keeping the amplitude spectrum unchanged, thereby stressing spatial organization under matched appearance statistics. Perturbations are applied only to the luminance channel in YCbCr color space to avoid chromatic artifacts. Additionally, a simple phase-saliency mask concentrates updates on the most influential frequencies. The model is trained with a weighted combination of losses on clean and phase-perturbed samples, supporting both single-source and multi-source DG. We validate our method on two challenging medical datasets and demonstrate that PhaseAT achieves over 20% improvement in single-source domain generalization, outperforming several state-of-the-art DG methods. The code implementation is available at: https://github.com/ahmed-sharshar/PhaseAT.